ROBOENV × RLE-BENCH · EVIDENCE REVIEW

Open the fridge.

Two agent harnesses. Five fresh episodes each. One shared evaluation protocol.

GPT-6 Astra / low · Standard

10 episodes · seeds 0–4 · Astra / low · standard tier. Kinex retains a 100-model-turn cap; Codex has no inhouse model-turn cap. This is an integration evaluation, not an isolated model-capability comparison. Read the full analysis · Browse all files · Data format · Episode CSV

Loading episode evidence…